Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS6944730 B2
Publication typeGrant
Application numberUS 10/352,090
Publication dateSep 13, 2005
Filing dateJan 28, 2003
Priority dateMar 1, 2002
Fee statusPaid
Also published asUS20030167385
Publication number10352090, 352090, US 6944730 B2, US 6944730B2, US-B2-6944730, US6944730 B2, US6944730B2
InventorsJiin Lai, Sheng-Chung Wu
Original AssigneeVia Technologies, Inc.
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Read/write scheduling apparatus of controller chip and method for the same
US 6944730 B2
Abstract
A read/write scheduling apparatus of controller chip and method for the same. The read/write scheduling apparatus is used for arbitrating a plurality of read and write requests from a CPU to access a memory unit. The read request has higher priority in a host bandwidth limited case and the write requests in write queues are not sent until a predetermined amount of write requests are accumulated. In a DRAM bandwidth limited case, the read and the write requests have the same priority. The scheduling apparatus counts the number of the read and write requests to the memory unit within a predetermined time, the operation is changed to DRAM bandwidth limited case in case that the counted number is larger than a predetermined number.
Images(3)
Previous page
Next page
Claims(14)
1. A read/write scheduling apparatus connected between a CPU and a memory unit, comprising:
a read queue connected to the CPU and temporarily storing a plurality of read requests issued from the CPU for accessing the memory unit;
a write queue connected to the CPU and temporarily storing a plurality of write requests issued from the CPU for accessing the memory unit;
a decision unit used discriminate an operation mode between a host bandwidth limited case and a memory bandwidth limited case according to the read and write requests from the CPU; and
an arbiter connected to the read queue and the write queue;
wherein in the memory bandwidth limited case, the arbiter continuously sending read and write requests stored in the read queue and the write queue to the memory unit.
2. The read/write scheduling apparatus as in claim 1, wherein the decision unit further comprises a setting timer used in the host bandwidth limited case, the setting timer counting the number of the read and write requests to the memory unit within a predetermined time, in case that the counted number is larger than a predetermined number, the timer will inform the arbiter to change the operation mode into the memory bandwidth limited case.
3. The read/write scheduling apparatus as in claim 1, wherein the decision unit further comprises a resetting timer used in the memory bandwidth limited case, the resetting timer counting the number of the read and write requests to the memory unit within a predetermined time, in case that the counted number is small than a predetermined number, the timer will inform the arbiter to change the operation mode into the host bandwidth limited case.
4. The read/write scheduling apparatus as in claim 1, wherein the arbiter sends the read requests to the memory unit with higher priority than the write requests in memory bandwidth limited case.
5. The read/write scheduling apparatus as in claim 1, wherein the memory unit is a DRAM memory unit.
6. The read/write scheduling apparatus as in claim 1, wherein the memory unit is a DDR DRAM memory unit.
7. The read/write scheduling apparatus as in claim 1, wherein the memory unit is a multiple data rate DRAM memory unit.
8. The read/write scheduling apparatus as in claim 1, wherein the memory unit is a Rambus memory unit.
9. A read/write scheduling method for arbitrating a plurality of read requests and a plurality of write requests issued from a CPU for accessing a memory unit, the method comprising following steps:
discriminating an operation mode between a host bandwidth limited case and a memory bandwidth limited case according to the read and write requests from the CPU;
in the memory bandwidth limited case, continuously sending the read and write requests stored in the read queue and the write queue to the memory unit, wherein the read requests have higher priority than the write request.
10. The read/write scheduling method as in claim 9, further comprising following steps:
counting the number of the read and write requests to the memory unit within a predetermined time in the host bandwidth limited case; and
changing the operation mode into the memory bandwidth limited case in case that the counted number is larger than a predetermined number.
11. The read/write scheduling method as in claim 9, further comprising following steps:
counting the number of the read and write requests to the memory unit within a predetermined time in the memory bandwidth limited case; and
changing the operation mode into the host bandwidth limited case in case that the counted number is not larger than a predetermined number.
12. A read/write scheduling method for arbitrating a plurality of read requests and a plurality of write requests issued from a CPU and used to access a memory unit, the method comprising following steps:
counting the number of the read and write requests to the memory unit within a first predetermined time;
changing an operation mode to a memory bandwidth limited case in case that the counted number is larger than a first predetermined number; and
continuously sending the read and write requests stored in the read queue and the write queue to the memory unit in the memory bandwidth limited case, wherein the read requests have higher priority than the write requests.
13. The read/write scheduling method as in claim 12, further comprising following step:
counting the number of the read and write requests to the memory unit within a second predetermined time in the memory bandwidth limited case;
changing operation mode into a host bandwidth limited case in case that the counted number is not larger than a second predetermined number.
14. The read/write scheduling method as in claim 12 further comprising following step:
waiting a plurality of clocks after the read requests are sent; and
sending the write requests to the memory unit after ensuring no read requests being present.
Description
FIELD OF THE INVENTION

The present invention relates to a read/write scheduling apparatus of controller chip and method for the same, especially to a read/write scheduling apparatus of controller chip for managing both the host bandwidth limited case and the DRAM bandwidth limited case of data accessing in computer system and method for the same.

BACKGROUND OF THE INVENTION

The CPU is the heart of a computer for executing command and other versatile function. To simplify layout, a controller chip is provided as interface between the CPU and memory (such as DRAM), accelerated graphic port (AGP) and other peripherals. The controller chip generally comprises FIFO (first-in first-out) queues to temporarily store the requests from the CPU and the controller chip manages the FIFO queues for efficiency operation of the computer system.

FIG. 1 shows a block diagram illustrating the arrangement of a prior art controller chip. The controller chip 6 is connected between a CPU 1 and a DRAM unit 2. The controller chip 6 is composed of a write queue 5, a read queue 4 and an arbiter 3. The write requests of the CPU 1 for accessing the DRAM unit 2 are firstly stored in the write queue 5 and the read requests of the CPU 1 for accessing the DRAM unit 2 are firstly stored in the read queue 4. The arbiter 3 is coupled between the write queue 5 (the read queue 4) and the DRAM unit 2 and controls the transmission of the requests stored in the write queue 5 (the read queue 4) to the DRAM unit 2 for regulating the transaction between the CPU 1 and the DRAM unit 2.

In conventional computer, the CPU 1 has a host bus with bandwidth less than the bandwidth of the memory bus. In case that the CPU 1 sends a plurality of requests to the DRAM unit 2, a host bandwidth limited case is encountered wherein the host bus bandwidth is congested and the memory bus bandwidth is sufficient. To exploit fully ability of the CPU 1, the scheduler in the arbiter 3 processes read requests with higher priority in host bandwidth limited case. The read requests are dominant such that the CPU 1 will rapidly fetch required data for efficient processing.

On the other hand, the write requests are firstly stored in the write queue 5 and the arbiter 3 will halt the write requests until the accumulated number of the write requests exceeds a predetermined value. Therefore, the overall efficiency of the computer system can be enhanced.

In present computer system, the host bus has bandwidth larger than the bandwidth of the memory bus. Therefore, the computer system is operated in a DRAM bandwidth limited case, especially when the CPU 1 has sent many requests to the DRAM unit 2 and the bus bandwidth of the DRAM unit 2 is occupied by the requests of the CPU 1. The scheduler in the arbiter 3 should have more sophisticated strategy to prevent efficiency degrade of the computer system.

SUMMARY OF THE INVENTION

In one aspect of the present invention, the present invention provides a read/write scheduling apparatus connected between a CPU and a memory unit and comprising a read queue connected to the CPU and temporarily storing a plurality of read requests issued from the CPU for accessing the memory unit; a write queue connected to the CPU and temporarily storing a plurality of write requests issued from the CPU for accessing the memory unit; a decision unit used discriminate between a host bandwidth limited case and a memory bandwidth limited case according to the read and write requests from the CPU; and an arbiter connected to the read queue and the write queue. In the memory bandwidth limited case, the arbiter continuously sends the read and write requests stored in the read queue and the write queue to the memory unit.

In another aspect of the present invention, the present invention provides a read/write scheduling method for arbitrating a plurality of read requests and a plurality of write requests issued from a CPU for accessing a memory unit. The method comprises following steps: discriminating between a host bandwidth limited case and a memory bandwidth limited case according to the read and write requests from the CPU; continuously sending the read and write requests stored in the read queue and the write queue to the memory unit in the memory bandwidth limited case, wherein the read requests have higher priority.

In still another aspect of the present invention, the present invention provides a read/write scheduling method for arbitrating a plurality of read requests and a plurality of write requests issued from a CPU for accessing a memory unit. The method comprises following steps: counting the number of the read and write requests to the memory unit within a first predetermined time, changing operation mode to a memory bandwidth limited case in case that the counted number is larger than a first predetermined number; and continuously sending the read and write requests stored in the read queue and the write queue to the memory unit in the memory bandwidth limited case, wherein the read requests have higher priority.

The various objects and advantages of the present invention will be more readily understood from the following detailed description when read in conjunction with the appended drawing, in which:

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 shows a block diagram illustrating the arrangement of a prior art controller chip; and

FIG. 2 shows a block diagram of the controller chip of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

The present invention is exemplified with a computer system with Pentium 4 CPU. The motherboard speed for the Pentium 4, CPU is 100 MHz, and the Pentium 4 CPU has 64 data lines and operates at 4× speed. Therefore, the host bus bandwidth is:
100 MHz×4×8 (byte)=3.2 Gbyte/s.

Moreover, the current bandwidth of double data rate (DDR) DRAM with 133 Hz frequency and 64 data lines is
133 MHz×2×8 (byte)=2.1 Gbyte/s.

In above-mentioned computer architecture, the host bus bandwidth is larger than the memory bus bandwidth. Therefore, the DRAM bandwidth limited case is encountered especially when the CPU 1 sends a plurality of read/write requests to the DRAM unit 2.

The important issue in the DRAM bandwidth limited case is to exploit the full capability of DRAM unit 2. The conventional arbiter halts the write queue until the accumulated number of the write requests exceeds a predetermined value, which is inefficient and degrades the performance of the computer system in DRAM bandwidth limited case.

FIG. 2 shows a block diagram of the controller chip 30 connected between a CPU 10 and a memory unit 20. The memory unit 20, for example, can be DRAM unit, DDR DRAM, multiple data rate RAM and Rambus memory. The controller chip 30 is used to schedule the read/write request of the CPU 10 to a DRAM unit 20 and comprises at least a decision unit 40, a write queue 60, a read queue 70 and an arbiter 80.

The decision unit 40 is used to judge whether the request from the CPU 10 to the DRAM unit 20 are issued in the host bandwidth limited case or the DRAM bandwidth-limited case. The computer system is in the host bandwidth limited case in case that the CPU 10 sends read/write request to the DRAM unit 20 in intermittent way. The arbiter 80 of the controller chip 30 works in similar way as conventional approach, wherein the write requests are firstly stored in the write queue 60 and will not be sent by the arbiter 80 until the accumulated number of the write requests exceeds a predetermined value.

On the other hand, the computer system is in the DRAM bandwidth limited case in case that the CPU 10 sends read/write request to the DRAM unit 20 in continuous way. In this case, the arbiter 80 of the controller chip 30 will continuously send read requests in the read queue 70 and the write requests in the write queue 60 to the DRAM unit 20 in order to fully exploit the capability of the DRAM unit 20. In other word, once the read/write request is present in the queue, the arbiter 80 of the controller chip 30 will send the request to the DRAM unit 20.

More particularly, in the DRAM bandwidth limited case, the DRAM unit 20 will operate in response to both read request and write request, while the CPU 10 will operate only to the read request. Accordingly, for the arbiter 80, the read request has higher priority than the write request when both read request and write request are present.

In the present invention, the decision unit 40 is realized by a timer. In host bandwidth limited case, a setting timer is activated to counts the number of the read/write requests to the DRAM unit 20 issued from the CPU 10 within a predetermined time. In case that the counted number is larger than a predetermined number, the setting timer will inform the arbiter 80 to change the operation mode into DRAM bandwidth limited case, and the arbiter 80 will deal the access request on DRAM bandwidth limited base as stated above. Moreover, in DRAM bandwidth limited case, a resetting timer is activated to counts the number of the read/write requests to the DRAM unit 20 issued from the CPU 10 within a predetermined time. In case that the counted number is smaller than a predetermined number, the resetting timer will inform the arbiter 80 to change the operation mode into host bandwidth limited case, and the arbiter 80 will deal the access request on host bandwidth limited base as stated above.

As can be seen from above description, the read/write scheduling apparatus of controller chip according to the present invention can enhance the data accessing performance of computer system; especially in DRAM bandwidth limited case.

Although the present invention has been described with reference to the preferred embodiment thereof, it will be understood that the invention is not limited to the details thereof. Various substitutions and modifications have suggested in the foregoing description, and other will occur to those of ordinary skill in the art. Therefore, all such substitutions and modifications are intended to be embraced within the scope of the invention as defined in the appended claims.

Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US5842038 *Oct 10, 1996Nov 24, 1998Unisys CorporationOptimized input/output memory access request system and method
US6088772 *Jun 13, 1997Jul 11, 2000Intel CorporationMethod and apparatus for improving system performance when reordering commands
US6092158 *Jun 13, 1997Jul 18, 2000Intel CorporationMethod and apparatus for arbitrating between command streams
US6295586 *Dec 4, 1998Sep 25, 2001Advanced Micro Devices, Inc.Queue based memory controller
US6496906 *Apr 30, 2001Dec 17, 2002Advanced Micro Devices, Inc.Queue based memory controller
US6564304 *Sep 1, 2000May 13, 2003Ati Technologies Inc.Memory processing system and method for accessing memory including reordering memory requests to reduce mode switching
US6601151 *Feb 8, 1999Jul 29, 2003Sun Microsystems, Inc.Apparatus and method for handling memory access requests in a data processing system
US20030070055 *Sep 28, 2001Apr 10, 2003Johnson Jerome J.Memory latency and bandwidth optimizations
US20030079073 *Sep 29, 2001Apr 24, 2003Richard Elizabeth AnneTransaction generator for initialization, rebuild, and verify of memory
Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US7596707 *May 6, 2005Sep 29, 2009Sun Microsystems, Inc.System and method for efficient power throttling in multiprocessor chip
US8001338Aug 21, 2007Aug 16, 2011Microsoft CorporationMulti-level DRAM controller to manage access to DRAM
US8180975 *May 15, 2012Microsoft CorporationControlling interference in shared memory systems using parallelism-aware batch scheduling
US8271741 *Nov 5, 2008Sep 18, 2012Microsoft CorporationPrioritization of multiple concurrent threads for scheduling requests to shared memory
US20090055580 *Aug 21, 2007Feb 26, 2009Microsoft CorporationMulti-level dram controller to manage access to dram
US20090216962 *Nov 5, 2008Aug 27, 2009Microsoft CorporationPrioritization of multiple concurrent threads for scheduling requests to shared memory
US20090217273 *Feb 26, 2008Aug 27, 2009Microsoft CorporationControlling interference in shared memory systems using parallelism-aware batch scheduling
Classifications
U.S. Classification711/158, 711/154, 711/167
International ClassificationG06F13/16
Cooperative ClassificationG06F13/161
European ClassificationG06F13/16A2
Legal Events
DateCodeEventDescription
Jan 28, 2003ASAssignment
Owner name: VIA TECHNOLOGIES, INC., TAIWAN
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:LAI, JIIN;WU, SHENG-CHUNG;REEL/FRAME:013709/0784
Effective date: 20020530
Mar 13, 2009FPAYFee payment
Year of fee payment: 4
Feb 13, 2013FPAYFee payment
Year of fee payment: 8